What this is
ORAILIX studies how modern AI systems, large language models in particular, succeed and fail at the requirements that make them deployable in regulated settings: correctness, confidentiality and accountability. The work is anchored by the Chair Trusted and Responsible AI at École Polytechnique, jointly funded by Crédit Agricole S.A.
The output is a set of auditable primitives, benchmarks, extraction attacks and interpretability tools, that make failure modes visible before deployment and defensible after.
Why it matters
Regulation (the EU AI Act, OECD AI Principles, sector guidance from banking, insurance and healthcare authorities) now requires operators to characterise the risks of their AI systems in reproducible terms. Yet most deployed models cannot be audited after the fact. Training data has been discarded, hallucinations are indistinguishable from valid outputs, and memorised private content leaks silently.
Representative work
- Guess or Recall? Training CNNs to classify and localise memorisation in LLMs (AAAI-26). Small classifiers detect and localise memorised content inside LLM outputs; memorisation is predictable from per-token signals, giving a tractable pre-deployment leakage check.
- MUCH, A Multilingual Claim Hallucination Benchmark. Measures claim-level hallucination across languages, surfacing systematic failure patterns that single-language evaluations miss.
- Unveiling Decision-Making in LLMs for Text Classification (EACL 2026). Sparse Autoencoders extract influential and interpretable concepts, giving regulators and auditors a route to post-hoc explanations.
- Reconstructing training data from document-understanding models (USENIX Security 2024). Exposed privacy weaknesses of production document pipelines; informed defences now being integrated with partners.
- Predicting memorisation within fine-tuned LLMs (ECAI 2025).
Partners
Industrial chair with Crédit Agricole S.A., collaborations with the Fondation de l'École Polytechnique and the TRAI master. Co-authors across the Institut Polytechnique de Paris ecosystem.
Related news
Follow related dispatches in the news stream, in particular the publications and talks channels.
← Back to ORAILIX